1. Field
The present invention relates to a driving assistance system including a prediction subsystem for passive or active assistance of a driver of a vehicle, and further relates to a corresponding method, software product, and a vehicle equipped with such a driving assistance system.
2. Description of Related Art
A manifold of driving assistance systems for vehicles is available today which aim at increasing driving comfort and/or safety of the passengers of a vehicle. Based on various sensor equipment such as radar, lidar, cameras, etc., functions related to driving or maneuvering range from distance sensing and parking assistance to sophisticated “Advanced Driver Assistant Systems” (ADAS) such as, for example, cruise-control functions, e.g. “Intelligent Adaptive Cruise Control” (IACC), which may include a lane change assistant, collision mitigation functions, emergency braking, etc.
Functions related to, e.g., ADAS may include a detection of other vehicles or objects moving in front or behind the ego-vehicle, and may include functions for predicting a future behavior of moving objects. The predictions may be used by the driving assistant for active control of the ego-vehicle and/or for providing information to the driver, e.g. for presenting a warning message via display or audio means to the driver.
While predictions serve generally well as a basis for decisions in advanced driver assistance systems, there remain problems. There is a general demand for advanced assistance functions to operate with high reliability, which includes avoiding situations that may let the driver feel uncomfortable or that may even require intervention of the driver.
However, sensor data are generally limited in accuracy due to limitations of the available sensor equipment, its range, orientation, etc., and due to inherent uncertainties of the measurement techniques. Besides error margins in the sensor data, sensor systems are also prone to errors such as misdetection, late detections, and/or wrong detections of entities in the environment of the host vehicle. These error sources may in turn lead to less reliable predictions, and driving assistants need to implement strategies to deal with uncertainties in the sensor data and sensing results.
On a more detailed level, the operation of some driving assistant functions may be based on sensing one entity only; as an example, a simple cruise control function may comprise keeping a predetermined distance to the preceding vehicle. In this case the operation is limited by the detection accuracy of the sensor data related to the detection of the preceding vehicle only. However, more sophisticated functions may require data related to multiple entities or objects, and may require even secondary data derived from the sensor data, such as data representing, e.g., a gap between two vehicles detected in the vicinity of the host vehicle. It is clear that an error such as a misdetection of one of the vehicles will cause an error in the assistant function; however, also mere inaccuracies in the detected positions of the two vehicles lead to an accumulated inaccuracy in the derived gap width which may result in a wrong decision on whether the gap is or will be sufficient for a third vehicle to perform a lane change. Predictions based on such decisions may in turn also be wrong and may result in a system response which appears confusing and unacceptable to the driver and/or other traffic participants.
The straightforward solution to enhancing system reliability is providing additional sensor equipment and/or high-performance equipment. This may serve to improve the available data basis, but at increasing hardware complexity and costs.
Assuming instead a given sensor equipment, various approaches to deal with sensor data inaccuracies are known for driving assistants with prediction subsystems. Some approaches explicitly assume perfect sensor equipment without taking further measures.
Broadhurst, A., et al., “Monte Carlo Road Safety Reasoning”, Intelligent Vehicles Symposium, 6-8 Jun. 2005, IEEE Proceedings 2005, p. 319-324, ISBN: 0-7803-8961-1, describe a framework for reasoning about the future motion of multiple objects in a road scene. Monte Carlo path planning is used to generate a probability distribution for the possible future motion of every car in the scene. The car may be controlled directly using the best predicted action, or the car may display a recommended path to the driver, or may display warnings on dangerous objects or regions on the road. Sensor uncertainty is said to be a future consideration.
According to another approach, errors in the perception of the environment are only implicitly considered.
US 2010/0228419 A1 describes a technique for risk assessment in an autonomic vehicle control system. Each of a plurality of objects detected proximate to a vehicle is monitored by various sensor equipment such as long- and short-range radar and a front camera. Sensor data are fused and, based on the fused data, object locations are predicted relative to a projected trajectory of the ego-vehicle. A collision risk level between the vehicle and each of the objects during a lane-change maneuver is assessed with respect to potential actions of the detected objects such as continuing with a fixed velocity, mild braking, or hard braking. A lane change maneuver is controlled according to the assessment and risk tolerance rules specifying spatial safety margins.
Sensor accuracy is discussed and it is appreciated that sensory detection and measurement of object locations and conditions are to be referred to as “estimates”. However, no explicit treatment of these estimates is performed any further. The fused object data comprise a degree of confidence in the data estimate.
EP 2 562 060 A1 (EP'060 for short hereinafter) describes a technique in a host vehicle for predicting a movement behavior of a target traffic object with exemplary emphasis on target objects cutting-in to a lane of the host vehicle or cutting-out from the lane of the host vehicle. The technique is based on two separate prediction modules, wherein a context based prediction (CBP) is related to a recognition of a movement behavior, i.e. a determination of “what” will happen, while a physical prediction (PP) is related to a determination of “how” a behavior will or may happen. The context based prediction relies on at least indirect indicators, while the physical prediction relies on direct indicators.
An indicator comprises a measurable variable conveying information about the future or ongoing behavior of a target vehicle and a confidence value indicating the true-state of the measurable variable. The confidence value is obtained by combining the sensor-confidence of all perceived scene elements which have been evaluated for the computation of the measurable variable, wherein the sensor confidence is a value for the reliability of the sensed information. Indicators can be combined with each other.
Direct indicators comprise observable variables, which are observable if and only if the behavior to be detected has started. For example, for predicting a lane-change, a set of direct indicators may comprise one or more of a lateral velocity, a lateral position relative to the lane, a changing orientation relative to the lane, and a changing orientation relative to other traffic participants.
Indirect indicators comprise observable variables, which are already observable before the predicted behavior has started. Indirect indicators may be defined as a set of indicators excluding direct indicators. For example, indirect indicators may relate to information about a relation between at least one traffic participant and one or more other traffic participants or static scene elements, such as an indicator indicating whether or not a fitting gap is available on a lane neighboring to the host-vehicle.
Other indirect indicators may relate to information about driver intentions, which may actively be communicated by the traffic participant whose behavior is to be predicted. Examples are intentions presumably indicated with a turning-signal, a braking-light, or information received via car-to-car-communication.
A set of potential trajectories may be computed for a target vehicle. By using the predicted movement behaviors from CBP, the set of relevant trajectories may be reduced. Matching a situation model against the history of perceived positional data in PP may help to further reduce the relevant trajectories.
More specifically, for predicting a target vehicle's future positions, in a first step, the probability for the target vehicle to perform one of a set of possible movement behaviors is estimated by the CBP. Some or all of these movement behaviors are validated by means of a PP. The purpose of the physical prediction is twofold: First, it validates the set of possible trajectories against a combination of the results of the CBP, the physical evidence, and vehicle relations. Second, it estimates the future position of each vehicle. In a final step a mismatch detection analyzes the consistency of the PP and the CBP. In case of mismatch, a fallback to the PP can be performed.
The context based prediction, physical prediction, and mismatch detection can be encapsulated in situation specific models and may be performed by different hardware units within the driver assistance system. Suited models fitting to the vehicle's environment can be activated or deactivated based on environment perception or self-localization.
Active control resulting from a wrong prediction based on sensor inaccuracy or sensing errors may need to be stopped and reversed when the target vehicle shows an unpredicted behavior or a behavior which has been predicted with low probability. The resultant control may seem inappropriate, confusing and not comfortable to the driver and/or other traffic participants. The assistance system described in EP'060 therefore intends to minimize wrong predictions as far as possible by means of the introduction of situation models and a mismatch detection, amongst others.
According to still another approach to enhance system reliability, sensor uncertainty is modeled and may then directly or indirectly influence the prediction result. Sensor uncertainties can be modeled, e.g., based on assumptions of sensor accuracy. The estimated uncertainties may then influence the prediction result.
Dagli, I., et al., “Cutting-in Vehicle Recognition for ACC Systems—Towards Feasible Situation Analysis Methodologies”, Intelligent Vehicles Symposium, 14-17 Jun. 2004, IEEE Proceedings 2004, p. 925-930, ISBN: 0-7803-8310-9, describe a cutting-in vehicle recognition functionality for ACC systems that utilizes a probabilistic model for situation analysis and prediction. In order to cope with low sensor data quality, sensor data filtering is combined with Kalman filters and situation analysis with probabilistic networks, in order that low quality sensor data is faded out in the decision process.